BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20191206T142452Z
LOCATION:Analog 1\, 2
DTSTART;TZID=Europe/Stockholm:20180625T133700
DTEND;TZID=Europe/Stockholm:20180625T134100
UID:isc_hpc_ISC High Performance 2018_sess103_phd111@linklings.com
SUMMARY:(PhD10) Quantification of Load Balancing related Performance Optim
 izations
DESCRIPTION:PhD Forum\n\n(PhD10) Quantification of Load Balancing related 
 Performance Optimizations\n\nHarlacher\n\nQuantification of load balancing
  related performance optimizations\n\nWhile extensive research in the area
  of load balancing MPI applications broad up a wide range of metrics to de
 scribe and differentiate load imbalance as well as numerous methods and al
 gorithms to improve load imbalances, the field still lacks criteria to spe
 cify the applications sensitivity to load imbalances and measures to quant
 ify the actual costs and benefits of load balancing efforts. \nEarlier per
 formance analysis of two simulation codes showed that the implemented data
  structures, tailored to the highly tuned computational routines, provide 
 little flexibility for load balancing (LB) and restrict the LB quality. Im
 plementation and testing of alternative approaches (data structures and LB
  strategies) is unreasonable considering size and complexity of the source
  code. Application developers are left alone with the assessment and decis
 ion of expensive redesign issues in order to optimize application runtime.
 \n\nThe central idea of this thesis is the development a software engineer
 ing tool that facilitates MPI application developers to study the effects 
 of alternative load balancing strategies and data structures for their pro
 blems without time-consuming code modifications to the original applicatio
 n or cumbersome analytical modeling. The aim is to simulate the applicatio
 n's behaviour and runtime based on abstract information of the application
 . This includes a description of the application pattern as well as the sp
 ecification of the computational and communicational load per work item. T
 he user represents the application schedule as a sequence of computation a
 nd communication phases. Each phase may have one or more weights associate
 d with it to characterize static or dynamic work load per work item respec
 tively. Computation load describes the time it takes to process a single i
 tem or element and communication load models the connectivity between item
 s as well as their amount of data exchange. This information is provided b
 y the user either from direct measurement within the application, utilizat
 ion of a performance measurement tool, a hypothetical estimate or a combin
 ation of the above. This abstract application description decouples the ma
 pping of items to processors from the original application and enables to 
 apply different partitionings. For a given partitioning, the simulator exe
 cutes the application schedule with its prescribed weights on the target a
 rchitecture and allows for direct performance analysis. \n\nOne applicatio
 n of this load balancing simulator is to apply different load distribution
 s (partitionings) to the two simulation codes and compare the resulting si
 mulation timings. This quantification of potential performance optimizatio
 n co-determines the decision for expensive software re-design. Further use
  cases target the multi-stage nature of scientific applications: the coupl
 ing of different solvers in APESmate and the handling of multi-level grids
  in APES. We investigate the effects of alternative data structures and pr
 ocessing strategies by modeling and simulation in our load balancing simul
 ator. \n\nThis work implies detailed studies to understand and model the c
 omplex interplay between data structures, application pattern, partitionin
 g and system parameters. Software developers need applicable means and mod
 els to early identify optimization potential in their codes and make infor
 med choices about adaptions to load balancing strategy and data structure.
 \n\nPasses: Conference Pass, Exhibition Pass, Communication Optimization, 
 MPI, Parallel Applications, Performance Analysis and Optimization, Scienti
 fic Software Development\n\nTag: Conference Pass, Exhibition Pass, Communi
 cation Optimization, MPI, Parallel Applications, Performance Analysis and 
 Optimization, Scientific Software Development
URL:https://2018.isc-program.com/presentation/?id=phd111&sess=sess103
END:VEVENT
END:VCALENDAR